Lately I’ve been thinking about artificial intelligence. Ever since OpenAI announced ChatGPT, the news and hype around AI is overwhelming. I thought I’d wait for things to settle down a bit, but it just keeps going. With stock prices soaring and people trying to jam AI into everything from pillows to toothbrushes it can feel like we’re in the early phases of a mania.
The serious analysis on the potential impact of AI is centered on the same theme: productivity. In short, AI tools are the next leg of the information technology revolution’s dramatic increase in productivity because they expand the types of tasks computers can do effectively. To get a better sense of where we’re headed, I decided to get my hands (metaphorically) dirty and see if I could hack something out while learning about how these new tools work under the hood.
Fast AI and Kaggle
First, I have to give a shout out to the folks at fast.ai for their amazing course practical deep learning for coders and the folks at Kaggle.com for their intro to python because, I am not a coder! If you want to dig in, I’d recommend starting there.
The problem
One of the challenges in driving progress is that public policy is incredibly complex. The analysis that goes into making public policy must be robust if it is going to account for all of the various stakeholders and social-economic cross-currents, so consequently the products produced by public policy experts are dense. Communicating those products to build a consensus for change is difficult, and susceptible to mis- and disinformation from the opponents of change. Advocates are forced to reduce the case for change to broadly appealing, simple messages that generally oversell the benefits. Opponents can easily find some distasteful nuance in the products, take it out of context, and blow it out of proportion. Complicating this communication further is the lack of time and attention most of us have in our days to think about these topics, and general skepticism that we’re not getting the full story; so we turn to our trusted intermediaries as a shortcut.
A solution?
More and more our trusted intermediaries are becoming digital platforms, but these have mostly been limited to lightly regulated consumer transactions like entertainment, food, and travel. Rather than advocates only relying on broadly appealing themes, what if the public could ask an AI chat-bot about what’s important to them?
Using an example from Anthropic’s cookbook on github, I created a chat-bot prototype that can answer questions about the analysis that went into Arlington’s missing middle / expanded housing choice zoning update. You can pass it a pdf and questions, and get an answer. I fed it the Missing Middle Housing Study Research Bulletin 2 and asked it how many more housing units does Arlington need:
and got this:
Feel free to copy my notebook on google colab and ask your own questions! With this type of capability, a chat-bot could replace or augment the Frequently Asked Questions (FAQ) type content that often goes along with these policy analyses. These could in turn be wired up to voice activated smart assistants and allow the community to interact with the content on their own terms. “Alexa! How many trees will we have to cut down because of this?” “Ok Google, what will the impact on the school budget be?” “Siri, how will this lower housing costs?” In fact, instead of assuming what questions the public might have, policy makers could actually find out, and adapt to the concerns of the community in real time.
Not quite yet
Don’t get too excited. This is, at the moment, a toy i.e. thousands of hours away from a production/public ready application. I’m not even sure it’s advisable to create a production ready application with today’s tech because generative AI has serious safety issues. I chose Anthropic because of their stated priority on AI safety, but I would caution that there is little in the way of safeguards preventing these companies from pivoting to prioritize profit, and no generally accepted measures for safety that would even allow you to do a comparison between products. Fortunately, some smart folks are working on this problem.
Also, these tools are not magic. To put them to work on many use cases, a lot of work needs to be done; in this case to load all of the context and train and fine tune the model so that it’s responses were accurate across a wide range of topics, not just a single pdf. To maintain this capability, the organization would need to have tools to allow the non-IT team to update, test, and monitor the chat-bot. That’s all assuming the infrastructure, cybersecurity, and website integration are auto-magically handled by the IT team.
So what
If you haven’t already, it’s time to get your feet wet. People are creating some very interesting capabilities with AI, and the uses are expanding by the minute. It is still, however, very early days. There is an inordinate amount of hype, and there will be many blow-ups and scandals, but you cannot afford to ignore these tools. Start looking for menial tasks around the office that could go faster with the AI tools that are already being offered by your existing technology providers. Start with things that are low risk to get your team familiar with the benefits and pitfalls of using these tools while the mistakes and lessons learned filter through the system. Once your confident you can manage them effectively, you can consider incorporating them into more critical operations.
Impulse Consulting 